Not long ago, The Wall Street Journal let “Claudius Sennet,” an AI agent, manage its vending machine under the watchful gaze of “Seymour Cash,” a bot. The model monitored inventory and sales and placed orders. It was an autonomous vending machine, a very simple business, certainly a lot simpler than managing a supply chain.

In a matter of days, the Journal reported that: 

Claudius had given away nearly all its inventory for free — including a PlayStation 5 it had been talked into buying for ‘marketing purposes.’ It ordered a live fish. It offered to buy stun guns, pepper spray, cigarettes and underwear. Profits collapsed.

There are many such tales of AI failures. PwC reports that only 12% of CEOs have seen cost and revenue benefits.

Supply chains are a lot more chaotic than an isolated, deterministic vending machine — and managing them is an art and science that has been honed over decades, even centuries. Organizations have learned, through expensive lessons, how to design, source, manufacture, and distribute products reliably, day after day. Many of those lessons can illuminate a path to success with AI (or serve as a refresher for the seasoned AI pro).

Before there is a product to move, there must be a product. And that starts with product design.

The entire process will fail if the product is poorly designed. Firstly, the product must meet customers’ needs or it is useless. But the best design may be a product consumers never knew they wanted. No one was clamoring for a smartphone until Apple conceived, designed, and built it. Secondly, the design should be manufacturable (feasible) and cost-effective to produce at scale.

For AI, the product is the use case: 

• Does it solve a worthwhile problem for businesses and end-users? 
• Is it aligned with your organization’s business strategy?
• How will value be created and measured? 
• Will it be reduced time to resolve customer issues, increased warehouse flow-through, or reduced product returns and repairs? 
• What is difficult to measure but might be impacted? (Customer experience, employee up-skilling, etc.)

Is there an 'iPhone' use case — an unprecedented but transformative use case, one that no one has asked for, because no one has thought of doing it that way? (The iPhone was not a better version of an existing product.)

• Finally, how feasible is the use case? 
• Is the required data accessible? 
• Is it easy to implement and to integrate into existing solutions? 
• Will it require changes to workflows and to other parts of the organization?

Like product design, the use case is fundamental to the project’s success. Make sure there is a clear explanation of the value expected, how it will be measured, and the full costs of achieving it.

Supply chains can’t run without supplies or raw materials. Mining companies extract ore. Apparel brands source fiber and textiles. Pharmaceutical companies source ingredients for their drugs and rigorously test them. In all cases, the quality of the raw materials that go into the products affects the final product’s quality and yield downstream. That’s why supply chains will sometimes use serialization (uniquely marking individual items) and traceability (tracking its journey) to ensure the integrity of the final product.

A recent outbreak of salmonella tied to jalapeños highlighted the importance of traceability. It enabled a US fast-food giant to quickly identify the contaminated ingredient and pull it from restaurants, while health agencies and regulators were still looking for the source.

For AI, data is the raw material. It needs to be high quality to ensure the models perform at the highest level. Getting the right data is key. If data is stale, irrelevant, biased, or unstructured, it compromises the model’s quality and the project’s outcome.

Data should accurately represent the problem being solved. Aim for diverse data that captures the nuances and richness of the problem, including outliers and errors. Ask, 'What have we left out, what edge cases and exceptions could arise?' As much as possible, keep the data current, auditable, traceable, and versioned. This will be valuable when things go wrong. Just as serialization and traceability in supply chains help identify and extract contaminated products from stores, AI traceability helps developers understand model behavior and identify bad data that drags down model performance.

We have our quality raw materials — the next step is to make the product. The manufacturing process takes the raw materials and converts them into components. Manufacturers use standard operating procedures, quality controls, and repeatable processes to help ensure a high-quality product. In high-value manufacturing sectors, such as semiconductors, process control and sophisticated testing are especially important to ensure the chips perform as designed.

Model development should employ similar processes, including reproducible pipelines, version control, and testing environments.

To ensure high-quality data, use automated systems such as AI pipelines to clean and process raw data into reliable, structured data. They make it easier to develop models and complex machine learning workflows. Continuous integration and continuous delivery or deployment (CI/CD) pipelines automate the tedious processes of building, testing, and deploying code changes. This reduces errors and increases the speed and reliability of your software releases.

Finally, to make sure models perform as intended and are secure, use monitoring frameworks (such as data quality gates and bias detectors) that track and evaluate performance. They will help detect data and model drift early, so you can fix them and protect performance and business value. For example, a cost-per-decision tracker could be used in high-frequency models for warehouses that could run thousands of times each shift across hundreds of sites. Token costs could escalate quickly if the model’s efficiency deteriorates, so you want to catch that early.

Of course, components and parts are of little use on their own — they need to be assembled into finished products. Consider the 30,000 parts that typically go into a car or truck. Individually, the parts are useless, but integrated into a single harmonious functional vehicle, you have a marvel of engineering that is reliable, capable of incredible performance, and built to operate flawlessly in the real world day after day. Testing is the final part of this process, with end-of-line testing checking parts and the vehicle for functionality, durability and compliance to specifications.

Similarly, an AI model in isolation is of little value. But when integrated into a well-designed functional system, a network-enabled platform powered by a unified data source and combined with an agentic layer, the model can work wonders. But rigorous testing is critical to the success of the deployment. Data, models, and the system should be checked for bias, errors, hallucinations, and explainability, as well as for compliance with ethical and responsible frameworks and regulation.

With a connected network of thousands of suppliers and carriers, agents can see and do more. Agents can see beyond the factory, vehicle, or warehouse, sense issues deep in the supplier network, suggest remedies, see opportunities for backhauls and continuous moves, and a lot more. Agents properly integrated into users’ workflows with transparency become 'teammates' who users grow to trust and rely on to solve real-world problems day in and day out.

The cycle is not complete until products reach customers. It’s only then that value is realized, and sellers are paid. In supply chain, distribution networks affect cost, efficiency, speed, and customer experience.

Models must be deployed cost-effectively and operate reliably and responsibly with minimal latency to deliver optimal value.

For AI pilots, carefully consider the full cost of deployment, including cost per inference, latency, reliability, and scalability. Poor performance in any of these areas can compromise the project’s value. Have clear answers to questions such as:

• Can it be batch processed or does it need to be real time? 
• What’s the cost per token? 
• Can it be scaled cost-effectively? 

Don’t forget the user experience — the customer (end-user) is king. A non-intuitive interface, confusing dashboard, and sluggish performance will undermine adoption and use, and drag down your return on investment (ROI).

By treating AI initiatives more holistically as a supply chain system, we put the focus where it belongs, not just on the model and agents, but on the entire system for delivering business impact. We don’t measure supply chains by an isolated benchmark, but by their ability to meet demand at the lowest landed cost. AI projects should be held to a similar standard, by their ability to deliver the most organizational value at the lowest total cost and risk.

In business, the best product doesn’t always win. The winner is the right product, with the right quality and size, in the right place at the right time and at the right price. It takes the entire supply chain to achieve that. The AI model that wins will be the one that delivers transformative results through an intuitive experience, reliably and responsibly, day after day.